Merge ../linus
[pandora-kernel.git] / arch / x86_64 / kernel / kprobes.c
1 /*
2  *  Kernel Probes (KProbes)
3  *  arch/x86_64/kernel/kprobes.c
4  *
5  * This program is free software; you can redistribute it and/or modify
6  * it under the terms of the GNU General Public License as published by
7  * the Free Software Foundation; either version 2 of the License, or
8  * (at your option) any later version.
9  *
10  * This program is distributed in the hope that it will be useful,
11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
13  * GNU General Public License for more details.
14  *
15  * You should have received a copy of the GNU General Public License
16  * along with this program; if not, write to the Free Software
17  * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
18  *
19  * Copyright (C) IBM Corporation, 2002, 2004
20  *
21  * 2002-Oct     Created by Vamsi Krishna S <vamsi_krishna@in.ibm.com> Kernel
22  *              Probes initial implementation ( includes contributions from
23  *              Rusty Russell).
24  * 2004-July    Suparna Bhattacharya <suparna@in.ibm.com> added jumper probes
25  *              interface to access function arguments.
26  * 2004-Oct     Jim Keniston <kenistoj@us.ibm.com> and Prasanna S Panchamukhi
27  *              <prasanna@in.ibm.com> adapted for x86_64
28  * 2005-Mar     Roland McGrath <roland@redhat.com>
29  *              Fixed to handle %rip-relative addressing mode correctly.
30  * 2005-May     Rusty Lynch <rusty.lynch@intel.com>
31  *              Added function return probes functionality
32  */
33
34 #include <linux/config.h>
35 #include <linux/kprobes.h>
36 #include <linux/ptrace.h>
37 #include <linux/string.h>
38 #include <linux/slab.h>
39 #include <linux/preempt.h>
40 #include <linux/module.h>
41
42 #include <asm/cacheflush.h>
43 #include <asm/pgtable.h>
44 #include <asm/kdebug.h>
45 #include <asm/uaccess.h>
46
47 void jprobe_return_end(void);
48 static void __kprobes arch_copy_kprobe(struct kprobe *p);
49
50 DEFINE_PER_CPU(struct kprobe *, current_kprobe) = NULL;
51 DEFINE_PER_CPU(struct kprobe_ctlblk, kprobe_ctlblk);
52
53 /*
54  * returns non-zero if opcode modifies the interrupt flag.
55  */
56 static inline int is_IF_modifier(kprobe_opcode_t *insn)
57 {
58         switch (*insn) {
59         case 0xfa:              /* cli */
60         case 0xfb:              /* sti */
61         case 0xcf:              /* iret/iretd */
62         case 0x9d:              /* popf/popfd */
63                 return 1;
64         }
65
66         if (*insn  >= 0x40 && *insn <= 0x4f && *++insn == 0xcf)
67                 return 1;
68         return 0;
69 }
70
71 int __kprobes arch_prepare_kprobe(struct kprobe *p)
72 {
73         /* insn: must be on special executable page on x86_64. */
74         p->ainsn.insn = get_insn_slot();
75         if (!p->ainsn.insn) {
76                 return -ENOMEM;
77         }
78         arch_copy_kprobe(p);
79         return 0;
80 }
81
82 /*
83  * Determine if the instruction uses the %rip-relative addressing mode.
84  * If it does, return the address of the 32-bit displacement word.
85  * If not, return null.
86  */
87 static inline s32 *is_riprel(u8 *insn)
88 {
89 #define W(row,b0,b1,b2,b3,b4,b5,b6,b7,b8,b9,ba,bb,bc,bd,be,bf)                \
90         (((b0##UL << 0x0)|(b1##UL << 0x1)|(b2##UL << 0x2)|(b3##UL << 0x3) |   \
91           (b4##UL << 0x4)|(b5##UL << 0x5)|(b6##UL << 0x6)|(b7##UL << 0x7) |   \
92           (b8##UL << 0x8)|(b9##UL << 0x9)|(ba##UL << 0xa)|(bb##UL << 0xb) |   \
93           (bc##UL << 0xc)|(bd##UL << 0xd)|(be##UL << 0xe)|(bf##UL << 0xf))    \
94          << (row % 64))
95         static const u64 onebyte_has_modrm[256 / 64] = {
96                 /*      0 1 2 3 4 5 6 7 8 9 a b c d e f         */
97                 /*      -------------------------------         */
98                 W(0x00, 1,1,1,1,0,0,0,0,1,1,1,1,0,0,0,0)| /* 00 */
99                 W(0x10, 1,1,1,1,0,0,0,0,1,1,1,1,0,0,0,0)| /* 10 */
100                 W(0x20, 1,1,1,1,0,0,0,0,1,1,1,1,0,0,0,0)| /* 20 */
101                 W(0x30, 1,1,1,1,0,0,0,0,1,1,1,1,0,0,0,0), /* 30 */
102                 W(0x40, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* 40 */
103                 W(0x50, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* 50 */
104                 W(0x60, 0,0,1,1,0,0,0,0,0,1,0,1,0,0,0,0)| /* 60 */
105                 W(0x70, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0), /* 70 */
106                 W(0x80, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 80 */
107                 W(0x90, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* 90 */
108                 W(0xa0, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* a0 */
109                 W(0xb0, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0), /* b0 */
110                 W(0xc0, 1,1,0,0,1,1,1,1,0,0,0,0,0,0,0,0)| /* c0 */
111                 W(0xd0, 1,1,1,1,0,0,0,0,1,1,1,1,1,1,1,1)| /* d0 */
112                 W(0xe0, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* e0 */
113                 W(0xf0, 0,0,0,0,0,0,1,1,0,0,0,0,0,0,1,1)  /* f0 */
114                 /*      -------------------------------         */
115                 /*      0 1 2 3 4 5 6 7 8 9 a b c d e f         */
116         };
117         static const u64 twobyte_has_modrm[256 / 64] = {
118                 /*      0 1 2 3 4 5 6 7 8 9 a b c d e f         */
119                 /*      -------------------------------         */
120                 W(0x00, 1,1,1,1,0,0,0,0,0,0,0,0,0,1,0,1)| /* 0f */
121                 W(0x10, 1,1,1,1,1,1,1,1,1,0,0,0,0,0,0,0)| /* 1f */
122                 W(0x20, 1,1,1,1,1,0,1,0,1,1,1,1,1,1,1,1)| /* 2f */
123                 W(0x30, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0), /* 3f */
124                 W(0x40, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 4f */
125                 W(0x50, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 5f */
126                 W(0x60, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 6f */
127                 W(0x70, 1,1,1,1,1,1,1,0,0,0,0,0,1,1,1,1), /* 7f */
128                 W(0x80, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* 8f */
129                 W(0x90, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 9f */
130                 W(0xa0, 0,0,0,1,1,1,1,1,0,0,0,1,1,1,1,1)| /* af */
131                 W(0xb0, 1,1,1,1,1,1,1,1,0,0,1,1,1,1,1,1), /* bf */
132                 W(0xc0, 1,1,1,1,1,1,1,1,0,0,0,0,0,0,0,0)| /* cf */
133                 W(0xd0, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* df */
134                 W(0xe0, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* ef */
135                 W(0xf0, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,0)  /* ff */
136                 /*      -------------------------------         */
137                 /*      0 1 2 3 4 5 6 7 8 9 a b c d e f         */
138         };
139 #undef  W
140         int need_modrm;
141
142         /* Skip legacy instruction prefixes.  */
143         while (1) {
144                 switch (*insn) {
145                 case 0x66:
146                 case 0x67:
147                 case 0x2e:
148                 case 0x3e:
149                 case 0x26:
150                 case 0x64:
151                 case 0x65:
152                 case 0x36:
153                 case 0xf0:
154                 case 0xf3:
155                 case 0xf2:
156                         ++insn;
157                         continue;
158                 }
159                 break;
160         }
161
162         /* Skip REX instruction prefix.  */
163         if ((*insn & 0xf0) == 0x40)
164                 ++insn;
165
166         if (*insn == 0x0f) {    /* Two-byte opcode.  */
167                 ++insn;
168                 need_modrm = test_bit(*insn, twobyte_has_modrm);
169         } else {                /* One-byte opcode.  */
170                 need_modrm = test_bit(*insn, onebyte_has_modrm);
171         }
172
173         if (need_modrm) {
174                 u8 modrm = *++insn;
175                 if ((modrm & 0xc7) == 0x05) { /* %rip+disp32 addressing mode */
176                         /* Displacement follows ModRM byte.  */
177                         return (s32 *) ++insn;
178                 }
179         }
180
181         /* No %rip-relative addressing mode here.  */
182         return NULL;
183 }
184
185 static void __kprobes arch_copy_kprobe(struct kprobe *p)
186 {
187         s32 *ripdisp;
188         memcpy(p->ainsn.insn, p->addr, MAX_INSN_SIZE);
189         ripdisp = is_riprel(p->ainsn.insn);
190         if (ripdisp) {
191                 /*
192                  * The copied instruction uses the %rip-relative
193                  * addressing mode.  Adjust the displacement for the
194                  * difference between the original location of this
195                  * instruction and the location of the copy that will
196                  * actually be run.  The tricky bit here is making sure
197                  * that the sign extension happens correctly in this
198                  * calculation, since we need a signed 32-bit result to
199                  * be sign-extended to 64 bits when it's added to the
200                  * %rip value and yield the same 64-bit result that the
201                  * sign-extension of the original signed 32-bit
202                  * displacement would have given.
203                  */
204                 s64 disp = (u8 *) p->addr + *ripdisp - (u8 *) p->ainsn.insn;
205                 BUG_ON((s64) (s32) disp != disp); /* Sanity check.  */
206                 *ripdisp = disp;
207         }
208         p->opcode = *p->addr;
209 }
210
211 void __kprobes arch_arm_kprobe(struct kprobe *p)
212 {
213         *p->addr = BREAKPOINT_INSTRUCTION;
214         flush_icache_range((unsigned long) p->addr,
215                            (unsigned long) p->addr + sizeof(kprobe_opcode_t));
216 }
217
218 void __kprobes arch_disarm_kprobe(struct kprobe *p)
219 {
220         *p->addr = p->opcode;
221         flush_icache_range((unsigned long) p->addr,
222                            (unsigned long) p->addr + sizeof(kprobe_opcode_t));
223 }
224
225 void __kprobes arch_remove_kprobe(struct kprobe *p)
226 {
227         mutex_lock(&kprobe_mutex);
228         free_insn_slot(p->ainsn.insn);
229         mutex_unlock(&kprobe_mutex);
230 }
231
232 static inline void save_previous_kprobe(struct kprobe_ctlblk *kcb)
233 {
234         kcb->prev_kprobe.kp = kprobe_running();
235         kcb->prev_kprobe.status = kcb->kprobe_status;
236         kcb->prev_kprobe.old_rflags = kcb->kprobe_old_rflags;
237         kcb->prev_kprobe.saved_rflags = kcb->kprobe_saved_rflags;
238 }
239
240 static inline void restore_previous_kprobe(struct kprobe_ctlblk *kcb)
241 {
242         __get_cpu_var(current_kprobe) = kcb->prev_kprobe.kp;
243         kcb->kprobe_status = kcb->prev_kprobe.status;
244         kcb->kprobe_old_rflags = kcb->prev_kprobe.old_rflags;
245         kcb->kprobe_saved_rflags = kcb->prev_kprobe.saved_rflags;
246 }
247
248 static inline void set_current_kprobe(struct kprobe *p, struct pt_regs *regs,
249                                 struct kprobe_ctlblk *kcb)
250 {
251         __get_cpu_var(current_kprobe) = p;
252         kcb->kprobe_saved_rflags = kcb->kprobe_old_rflags
253                 = (regs->eflags & (TF_MASK | IF_MASK));
254         if (is_IF_modifier(p->ainsn.insn))
255                 kcb->kprobe_saved_rflags &= ~IF_MASK;
256 }
257
258 static void __kprobes prepare_singlestep(struct kprobe *p, struct pt_regs *regs)
259 {
260         regs->eflags |= TF_MASK;
261         regs->eflags &= ~IF_MASK;
262         /*single step inline if the instruction is an int3*/
263         if (p->opcode == BREAKPOINT_INSTRUCTION)
264                 regs->rip = (unsigned long)p->addr;
265         else
266                 regs->rip = (unsigned long)p->ainsn.insn;
267 }
268
269 /* Called with kretprobe_lock held */
270 void __kprobes arch_prepare_kretprobe(struct kretprobe *rp,
271                                       struct pt_regs *regs)
272 {
273         unsigned long *sara = (unsigned long *)regs->rsp;
274         struct kretprobe_instance *ri;
275
276         if ((ri = get_free_rp_inst(rp)) != NULL) {
277                 ri->rp = rp;
278                 ri->task = current;
279                 ri->ret_addr = (kprobe_opcode_t *) *sara;
280
281                 /* Replace the return addr with trampoline addr */
282                 *sara = (unsigned long) &kretprobe_trampoline;
283
284                 add_rp_inst(ri);
285         } else {
286                 rp->nmissed++;
287         }
288 }
289
290 int __kprobes kprobe_handler(struct pt_regs *regs)
291 {
292         struct kprobe *p;
293         int ret = 0;
294         kprobe_opcode_t *addr = (kprobe_opcode_t *)(regs->rip - sizeof(kprobe_opcode_t));
295         struct kprobe_ctlblk *kcb;
296
297         /*
298          * We don't want to be preempted for the entire
299          * duration of kprobe processing
300          */
301         preempt_disable();
302         kcb = get_kprobe_ctlblk();
303
304         /* Check we're not actually recursing */
305         if (kprobe_running()) {
306                 p = get_kprobe(addr);
307                 if (p) {
308                         if (kcb->kprobe_status == KPROBE_HIT_SS &&
309                                 *p->ainsn.insn == BREAKPOINT_INSTRUCTION) {
310                                 regs->eflags &= ~TF_MASK;
311                                 regs->eflags |= kcb->kprobe_saved_rflags;
312                                 goto no_kprobe;
313                         } else if (kcb->kprobe_status == KPROBE_HIT_SSDONE) {
314                                 /* TODO: Provide re-entrancy from
315                                  * post_kprobes_handler() and avoid exception
316                                  * stack corruption while single-stepping on
317                                  * the instruction of the new probe.
318                                  */
319                                 arch_disarm_kprobe(p);
320                                 regs->rip = (unsigned long)p->addr;
321                                 reset_current_kprobe();
322                                 ret = 1;
323                         } else {
324                                 /* We have reentered the kprobe_handler(), since
325                                  * another probe was hit while within the
326                                  * handler. We here save the original kprobe
327                                  * variables and just single step on instruction
328                                  * of the new probe without calling any user
329                                  * handlers.
330                                  */
331                                 save_previous_kprobe(kcb);
332                                 set_current_kprobe(p, regs, kcb);
333                                 kprobes_inc_nmissed_count(p);
334                                 prepare_singlestep(p, regs);
335                                 kcb->kprobe_status = KPROBE_REENTER;
336                                 return 1;
337                         }
338                 } else {
339                         if (*addr != BREAKPOINT_INSTRUCTION) {
340                         /* The breakpoint instruction was removed by
341                          * another cpu right after we hit, no further
342                          * handling of this interrupt is appropriate
343                          */
344                                 regs->rip = (unsigned long)addr;
345                                 ret = 1;
346                                 goto no_kprobe;
347                         }
348                         p = __get_cpu_var(current_kprobe);
349                         if (p->break_handler && p->break_handler(p, regs)) {
350                                 goto ss_probe;
351                         }
352                 }
353                 goto no_kprobe;
354         }
355
356         p = get_kprobe(addr);
357         if (!p) {
358                 if (*addr != BREAKPOINT_INSTRUCTION) {
359                         /*
360                          * The breakpoint instruction was removed right
361                          * after we hit it.  Another cpu has removed
362                          * either a probepoint or a debugger breakpoint
363                          * at this address.  In either case, no further
364                          * handling of this interrupt is appropriate.
365                          * Back up over the (now missing) int3 and run
366                          * the original instruction.
367                          */
368                         regs->rip = (unsigned long)addr;
369                         ret = 1;
370                 }
371                 /* Not one of ours: let kernel handle it */
372                 goto no_kprobe;
373         }
374
375         set_current_kprobe(p, regs, kcb);
376         kcb->kprobe_status = KPROBE_HIT_ACTIVE;
377
378         if (p->pre_handler && p->pre_handler(p, regs))
379                 /* handler has already set things up, so skip ss setup */
380                 return 1;
381
382 ss_probe:
383         prepare_singlestep(p, regs);
384         kcb->kprobe_status = KPROBE_HIT_SS;
385         return 1;
386
387 no_kprobe:
388         preempt_enable_no_resched();
389         return ret;
390 }
391
392 /*
393  * For function-return probes, init_kprobes() establishes a probepoint
394  * here. When a retprobed function returns, this probe is hit and
395  * trampoline_probe_handler() runs, calling the kretprobe's handler.
396  */
397  void kretprobe_trampoline_holder(void)
398  {
399         asm volatile (  ".global kretprobe_trampoline\n"
400                         "kretprobe_trampoline: \n"
401                         "nop\n");
402  }
403
404 /*
405  * Called when we hit the probe point at kretprobe_trampoline
406  */
407 int __kprobes trampoline_probe_handler(struct kprobe *p, struct pt_regs *regs)
408 {
409         struct kretprobe_instance *ri = NULL;
410         struct hlist_head *head;
411         struct hlist_node *node, *tmp;
412         unsigned long flags, orig_ret_address = 0;
413         unsigned long trampoline_address =(unsigned long)&kretprobe_trampoline;
414
415         spin_lock_irqsave(&kretprobe_lock, flags);
416         head = kretprobe_inst_table_head(current);
417
418         /*
419          * It is possible to have multiple instances associated with a given
420          * task either because an multiple functions in the call path
421          * have a return probe installed on them, and/or more then one return
422          * return probe was registered for a target function.
423          *
424          * We can handle this because:
425          *     - instances are always inserted at the head of the list
426          *     - when multiple return probes are registered for the same
427          *       function, the first instance's ret_addr will point to the
428          *       real return address, and all the rest will point to
429          *       kretprobe_trampoline
430          */
431         hlist_for_each_entry_safe(ri, node, tmp, head, hlist) {
432                 if (ri->task != current)
433                         /* another task is sharing our hash bucket */
434                         continue;
435
436                 if (ri->rp && ri->rp->handler)
437                         ri->rp->handler(ri, regs);
438
439                 orig_ret_address = (unsigned long)ri->ret_addr;
440                 recycle_rp_inst(ri);
441
442                 if (orig_ret_address != trampoline_address)
443                         /*
444                          * This is the real return address. Any other
445                          * instances associated with this task are for
446                          * other calls deeper on the call stack
447                          */
448                         break;
449         }
450
451         BUG_ON(!orig_ret_address || (orig_ret_address == trampoline_address));
452         regs->rip = orig_ret_address;
453
454         reset_current_kprobe();
455         spin_unlock_irqrestore(&kretprobe_lock, flags);
456         preempt_enable_no_resched();
457
458         /*
459          * By returning a non-zero value, we are telling
460          * kprobe_handler() that we don't want the post_handler
461          * to run (and have re-enabled preemption)
462          */
463         return 1;
464 }
465
466 /*
467  * Called after single-stepping.  p->addr is the address of the
468  * instruction whose first byte has been replaced by the "int 3"
469  * instruction.  To avoid the SMP problems that can occur when we
470  * temporarily put back the original opcode to single-step, we
471  * single-stepped a copy of the instruction.  The address of this
472  * copy is p->ainsn.insn.
473  *
474  * This function prepares to return from the post-single-step
475  * interrupt.  We have to fix up the stack as follows:
476  *
477  * 0) Except in the case of absolute or indirect jump or call instructions,
478  * the new rip is relative to the copied instruction.  We need to make
479  * it relative to the original instruction.
480  *
481  * 1) If the single-stepped instruction was pushfl, then the TF and IF
482  * flags are set in the just-pushed eflags, and may need to be cleared.
483  *
484  * 2) If the single-stepped instruction was a call, the return address
485  * that is atop the stack is the address following the copied instruction.
486  * We need to make it the address following the original instruction.
487  */
488 static void __kprobes resume_execution(struct kprobe *p,
489                 struct pt_regs *regs, struct kprobe_ctlblk *kcb)
490 {
491         unsigned long *tos = (unsigned long *)regs->rsp;
492         unsigned long next_rip = 0;
493         unsigned long copy_rip = (unsigned long)p->ainsn.insn;
494         unsigned long orig_rip = (unsigned long)p->addr;
495         kprobe_opcode_t *insn = p->ainsn.insn;
496
497         /*skip the REX prefix*/
498         if (*insn >= 0x40 && *insn <= 0x4f)
499                 insn++;
500
501         switch (*insn) {
502         case 0x9c:              /* pushfl */
503                 *tos &= ~(TF_MASK | IF_MASK);
504                 *tos |= kcb->kprobe_old_rflags;
505                 break;
506         case 0xc3:              /* ret/lret */
507         case 0xcb:
508         case 0xc2:
509         case 0xca:
510                 regs->eflags &= ~TF_MASK;
511                 /* rip is already adjusted, no more changes required*/
512                 return;
513         case 0xe8:              /* call relative - Fix return addr */
514                 *tos = orig_rip + (*tos - copy_rip);
515                 break;
516         case 0xff:
517                 if ((*insn & 0x30) == 0x10) {
518                         /* call absolute, indirect */
519                         /* Fix return addr; rip is correct. */
520                         next_rip = regs->rip;
521                         *tos = orig_rip + (*tos - copy_rip);
522                 } else if (((*insn & 0x31) == 0x20) ||  /* jmp near, absolute indirect */
523                            ((*insn & 0x31) == 0x21)) {  /* jmp far, absolute indirect */
524                         /* rip is correct. */
525                         next_rip = regs->rip;
526                 }
527                 break;
528         case 0xea:              /* jmp absolute -- rip is correct */
529                 next_rip = regs->rip;
530                 break;
531         default:
532                 break;
533         }
534
535         regs->eflags &= ~TF_MASK;
536         if (next_rip) {
537                 regs->rip = next_rip;
538         } else {
539                 regs->rip = orig_rip + (regs->rip - copy_rip);
540         }
541 }
542
543 int __kprobes post_kprobe_handler(struct pt_regs *regs)
544 {
545         struct kprobe *cur = kprobe_running();
546         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
547
548         if (!cur)
549                 return 0;
550
551         if ((kcb->kprobe_status != KPROBE_REENTER) && cur->post_handler) {
552                 kcb->kprobe_status = KPROBE_HIT_SSDONE;
553                 cur->post_handler(cur, regs, 0);
554         }
555
556         resume_execution(cur, regs, kcb);
557         regs->eflags |= kcb->kprobe_saved_rflags;
558
559         /* Restore the original saved kprobes variables and continue. */
560         if (kcb->kprobe_status == KPROBE_REENTER) {
561                 restore_previous_kprobe(kcb);
562                 goto out;
563         }
564         reset_current_kprobe();
565 out:
566         preempt_enable_no_resched();
567
568         /*
569          * if somebody else is singlestepping across a probe point, eflags
570          * will have TF set, in which case, continue the remaining processing
571          * of do_debug, as if this is not a probe hit.
572          */
573         if (regs->eflags & TF_MASK)
574                 return 0;
575
576         return 1;
577 }
578
579 int __kprobes kprobe_fault_handler(struct pt_regs *regs, int trapnr)
580 {
581         struct kprobe *cur = kprobe_running();
582         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
583         const struct exception_table_entry *fixup;
584
585         switch(kcb->kprobe_status) {
586         case KPROBE_HIT_SS:
587         case KPROBE_REENTER:
588                 /*
589                  * We are here because the instruction being single
590                  * stepped caused a page fault. We reset the current
591                  * kprobe and the rip points back to the probe address
592                  * and allow the page fault handler to continue as a
593                  * normal page fault.
594                  */
595                 regs->rip = (unsigned long)cur->addr;
596                 regs->eflags |= kcb->kprobe_old_rflags;
597                 if (kcb->kprobe_status == KPROBE_REENTER)
598                         restore_previous_kprobe(kcb);
599                 else
600                         reset_current_kprobe();
601                 preempt_enable_no_resched();
602                 break;
603         case KPROBE_HIT_ACTIVE:
604         case KPROBE_HIT_SSDONE:
605                 /*
606                  * We increment the nmissed count for accounting,
607                  * we can also use npre/npostfault count for accouting
608                  * these specific fault cases.
609                  */
610                 kprobes_inc_nmissed_count(cur);
611
612                 /*
613                  * We come here because instructions in the pre/post
614                  * handler caused the page_fault, this could happen
615                  * if handler tries to access user space by
616                  * copy_from_user(), get_user() etc. Let the
617                  * user-specified handler try to fix it first.
618                  */
619                 if (cur->fault_handler && cur->fault_handler(cur, regs, trapnr))
620                         return 1;
621
622                 /*
623                  * In case the user-specified fault handler returned
624                  * zero, try to fix up.
625                  */
626                 fixup = search_exception_tables(regs->rip);
627                 if (fixup) {
628                         regs->rip = fixup->fixup;
629                         return 1;
630                 }
631
632                 /*
633                  * fixup() could not handle it,
634                  * Let do_page_fault() fix it.
635                  */
636                 break;
637         default:
638                 break;
639         }
640         return 0;
641 }
642
643 /*
644  * Wrapper routine for handling exceptions.
645  */
646 int __kprobes kprobe_exceptions_notify(struct notifier_block *self,
647                                        unsigned long val, void *data)
648 {
649         struct die_args *args = (struct die_args *)data;
650         int ret = NOTIFY_DONE;
651
652         if (args->regs && user_mode(args->regs))
653                 return ret;
654
655         switch (val) {
656         case DIE_INT3:
657                 if (kprobe_handler(args->regs))
658                         ret = NOTIFY_STOP;
659                 break;
660         case DIE_DEBUG:
661                 if (post_kprobe_handler(args->regs))
662                         ret = NOTIFY_STOP;
663                 break;
664         case DIE_GPF:
665         case DIE_PAGE_FAULT:
666                 /* kprobe_running() needs smp_processor_id() */
667                 preempt_disable();
668                 if (kprobe_running() &&
669                     kprobe_fault_handler(args->regs, args->trapnr))
670                         ret = NOTIFY_STOP;
671                 preempt_enable();
672                 break;
673         default:
674                 break;
675         }
676         return ret;
677 }
678
679 int __kprobes setjmp_pre_handler(struct kprobe *p, struct pt_regs *regs)
680 {
681         struct jprobe *jp = container_of(p, struct jprobe, kp);
682         unsigned long addr;
683         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
684
685         kcb->jprobe_saved_regs = *regs;
686         kcb->jprobe_saved_rsp = (long *) regs->rsp;
687         addr = (unsigned long)(kcb->jprobe_saved_rsp);
688         /*
689          * As Linus pointed out, gcc assumes that the callee
690          * owns the argument space and could overwrite it, e.g.
691          * tailcall optimization. So, to be absolutely safe
692          * we also save and restore enough stack bytes to cover
693          * the argument area.
694          */
695         memcpy(kcb->jprobes_stack, (kprobe_opcode_t *)addr,
696                         MIN_STACK_SIZE(addr));
697         regs->eflags &= ~IF_MASK;
698         regs->rip = (unsigned long)(jp->entry);
699         return 1;
700 }
701
702 void __kprobes jprobe_return(void)
703 {
704         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
705
706         asm volatile ("       xchg   %%rbx,%%rsp     \n"
707                       "       int3                      \n"
708                       "       .globl jprobe_return_end  \n"
709                       "       jprobe_return_end:        \n"
710                       "       nop                       \n"::"b"
711                       (kcb->jprobe_saved_rsp):"memory");
712 }
713
714 int __kprobes longjmp_break_handler(struct kprobe *p, struct pt_regs *regs)
715 {
716         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
717         u8 *addr = (u8 *) (regs->rip - 1);
718         unsigned long stack_addr = (unsigned long)(kcb->jprobe_saved_rsp);
719         struct jprobe *jp = container_of(p, struct jprobe, kp);
720
721         if ((addr > (u8 *) jprobe_return) && (addr < (u8 *) jprobe_return_end)) {
722                 if ((long *)regs->rsp != kcb->jprobe_saved_rsp) {
723                         struct pt_regs *saved_regs =
724                             container_of(kcb->jprobe_saved_rsp,
725                                             struct pt_regs, rsp);
726                         printk("current rsp %p does not match saved rsp %p\n",
727                                (long *)regs->rsp, kcb->jprobe_saved_rsp);
728                         printk("Saved registers for jprobe %p\n", jp);
729                         show_registers(saved_regs);
730                         printk("Current registers\n");
731                         show_registers(regs);
732                         BUG();
733                 }
734                 *regs = kcb->jprobe_saved_regs;
735                 memcpy((kprobe_opcode_t *) stack_addr, kcb->jprobes_stack,
736                        MIN_STACK_SIZE(stack_addr));
737                 preempt_enable_no_resched();
738                 return 1;
739         }
740         return 0;
741 }
742
743 static struct kprobe trampoline_p = {
744         .addr = (kprobe_opcode_t *) &kretprobe_trampoline,
745         .pre_handler = trampoline_probe_handler
746 };
747
748 int __init arch_init_kprobes(void)
749 {
750         return register_kprobe(&trampoline_p);
751 }